A cnd / cvi digital model and prediction method

Abstract

The application discloses a CVD / CVI digital model and prediction method, which comprises a first H2 barrel, an Ar barrel, a second H2 barrel, a chemical vapor deposition reactor, a gas mixer and an electronic balance with a heating function; the output port of the first H2 barrel, the output port of the Ar barrel and the electronic balance with the heating function are all connected with the gas mixer; the output port of the second H2 barrel is connected with the gas mixer; and the gas mixer is connected with the gas inlet of the chemical vapor deposition reactor; the application realizes fast and accurate prediction of the deposition rate of SiC in a wide CVD / CVI process parameter range, and lays a theoretical and method foundation for future CVD / CVI digital twinning and intelligent manufacturing.

Description

Technical Field

[0001] This invention relates to the field of chemical vapor phase (CVD), and in particular to a digital model and prediction method for CVD / CVI. Background Technology

[0002] Ceramic materials possess excellent properties such as high temperature resistance, high specific strength, high specific modulus, corrosion resistance, wear resistance, and low density, making them potential replacements for metals as a new generation of high-temperature structural materials. However, brittleness is a fatal weakness of ceramic materials, with a critical crack length of only tens of micrometers, thus they lack the plastic deformation capacity of metals.

[0003] Improving the brittleness of ceramic materials has been a long-standing concern for ceramic researchers. Among the various toughening approaches developed, continuous fiber-reinforced ceramic matrix composites (CFR–CMC, or CMC for short) can fundamentally overcome the brittleness of ceramics. They are indispensable materials for the development of high-tech fields such as aerospace and have become the mainstream direction for the development of ceramic matrix composites.

[0004] Among them, continuous fiber-reinforced silicon carbide ceramic matrix composites (CMC–SiC) are the most studied and most successfully applied new thermal structural ceramic matrix composites, with a series of outstanding advantages such as high temperature resistance, low density, high specific strength, high specific modulus, ablation resistance, oxidation resistance, and the ability to avoid catastrophic damage.

[0005] CMC–SiC mainly includes two types: carbon fiber-toughened silicon carbide (C / SiC) and silicon carbide fiber-toughened silicon carbide (SiC / SiC). Due to the low cost and easy availability of carbon fiber, C / SiC has become the preferred choice for the research, evaluation, and application of SiC ceramic matrix composites. It has been successfully used in the thermal structures of third-generation aero-engines, ramjet engines, solid and liquid engines, as well as in the thermal protection structures of aerospace vehicles, and has broad application prospects in the aviation and aerospace fields.

[0006] Chemical vapor deposition (CVD) and chemical vapor infiltration (CVI) are currently the most widely used and commercially available methods for preparing continuous fiber-reinforced silicon carbide ceramic matrix composites (CMC-SiC). However, due to the complex relationship between the deposition reaction and gas-phase mass transfer, CVD processes are difficult to control and have long processing cycles. Computer numerical simulation helps to understand the CVI process mechanism, predict the densification process, effectively analyze experimental data, and shorten the cycle of process parameter optimization. Therefore, it has significant positive implications and is widely considered an important supplement to the CVD / CVI experimental process.

[0007] With the rapid development of computer technology and interdisciplinary research methods, the concept of digital twins has led to the creation of physical model twins that correspond to actual production to achieve intelligent manufacturing and production. This new model of digital manufacturing brings new ideas to the research and development of CVD / CVI.

[0008] A crucial aspect of digital twins is numerical simulation, a vital step in establishing a correspondence between the actual reaction equipment and conditions. Unlike traditional simulation, digital twin technology adds a digital concept to traditional numerical simulation. This invention uses CVD-SiC as an example to illustrate how to collect a set of experimental data to establish a corresponding physical simulation model. Then, by combining new offline or online experimental data, a process database and machine model are established to optimize process parameters, thereby improving R&D and production efficiency and laying the theoretical and methodological foundation for future CVD / CVI digital twins and intelligent manufacturing.

[0009] Technical solution of existing technology 1

[0010] In recent years, researchers have conducted extensive studies on CVD / CVI using chemical engineering theory and computational fluid dynamics simulation techniques. Based on the geometric characteristics of the experimental reactor, a geometric solid model is created, and models of fluid flow, heat and mass transfer, and chemical reactions are established. Model parameters are defined, boundary conditions and region conditions are set, the model is divided into mesh cells, and the equations are solved using self-developed programs or commercial fluid dynamics software. Model parameters are estimated through comparison with pre-experimental results and parametric analysis. Finally, the established model is used to optimize the process design.

[0011] Disadvantages of existing technology 1

[0012] There are too many empirical parameters, and their universality is poor.

[0013] Technical solution of existing technology 2

[0014] Machine learning (ML) is being used more and more widely in CVD / CVI, but it is mainly applied to material morphology design and process parameter optimization.

[0015] Disadvantages of existing technology 2

[0016] No researchers have yet applied machine learning (ML) to predict SiC-CVD deposition rates, nor have they correlated it with process parameters. Furthermore, it requires large amounts of training data, resulting in high data costs. It is also highly dependent on data distribution and has poor predictive ability in areas with sparse data. Summary of the Invention

[0017] The purpose of this invention is to solve the problems existing in the prior art and provide a CVD / CVI digital model and prediction method. This invention enables rapid and accurate prediction of SiC deposition rate within a wide range of CVD / CVI process parameters, laying a theoretical and methodological foundation for future CVD / CVI digital twins and intelligent manufacturing.

[0018] The specific technical solution is as follows:

[0019] A CVD / CVI digital model includes a first H2 tank, an Ar tank, a second H2 tank, a chemical vapor deposition reactor, a gas mixer, and an electronic balance with heating function. The output ports of the first H2 tank, the Ar tank, and the electronic balance with heating function are all connected to the gas mixer. The output port of the second H2 tank is connected to the gas mixer, and the gas mixer is connected to the inlet of the chemical vapor deposition reactor.

[0020] Preferably, a graphite substrate is provided on one side of the chemical vapor deposition reactor, a graphite heater is provided on the other side of the chemical vapor deposition reactor, a porous disk is provided at the front end of the cross-section of the inside of the chemical vapor deposition reactor, and an exhaust port is provided at the bottom cross-section of the chemical vapor deposition reactor.

[0021] Preferably, the inlet diameter of the chemical vapor deposition reactor is 16 mm, the length of the reaction zone of the chemical vapor deposition reactor is 200 mm or 400 mm, the diameter of the chemical vapor deposition reactor is 25 mm, and the thickness is 6.5 mm.

[0022] The inlet diameter of the chemical vapor deposition reactor is 16 mm, and the length of the reaction zone of the chemical vapor deposition reactor includes 200 mm or 400 mm; the surface diameter of the chemical vapor deposition reactor is 25 mm, and the thickness is 6.5 mm.

[0023] A predictive method for CVD / CVI digital modeling includes the following steps:

[0024] Step S1: Using the MTS(CH3SiCl3)-Ar-H2 system, SiC coating or SiC substrate is prepared on a graphite substrate by chemical vapor infiltration to generate SiC;

[0025] Step S2: Establish a two-dimensional axisymmetric geometric model for the CVD / CVI chemical vapor deposition reactor;

[0026] Step S3: The two-dimensional axisymmetric geometric model of the CVD / CVI chemical vapor deposition reactor was meshed using finite element software, and a mesh model of the reactor with a reaction zone diameter of 100 mm and the chemical reaction and its kinetic parameters were constructed.

[0027] Step S4: Process the existing experimental data for parameter A in the chemical reaction and its kinetic parameters using an algorithm to finally obtain the value of parameter A;

[0028] Step S5: Assess the A parameter values ​​and data distribution within the training dataset to form a mixed dataset;

[0029] Step S6: Train the ML model based on the mixed dataset;

[0030] Step S7: For the trained hybrid dataset, the XG Boost algorithm and the Random forest algorithm are used for training, and then combined with machine learning algorithms to predict the CVD / CVI deposition rate.

[0031] Preferably, step S1 includes the following sub-steps:

[0032] Sub-step S11: The flow rates of H2 and Ar are controlled by a mass flow controller;

[0033] Sub-step S12: By heating the MTS solution, H2 bubbles in the MTS solution and carries out the MTS.

[0034] Sub-step S13: After thorough mixing in all gas mixers, the mixture is piped into the chemical vapor deposition reactor;

[0035] Sub-step S14: The reaction gas passes through a porous gas distribution plate and is gradually heated. During transport, a gas phase reaction occurs and SiC is adsorbed on the graphite substrate. The tail gas is discharged from the outlet.

[0036] Preferably, the algorithm processing in step S4 includes the following sub-steps:

[0037] Sub-step S41: The A parameter needs to be fitted based on the existing experimental data, and combined with the model, the "finding the optimal model parameter value" is transformed into a "function extremum optimization problem";

[0038] Sub-step S42: Input the genetic algorithm, which optimizes the elements of the BP neural network and finally obtains the A parameter value.

[0039] The elements of optimizing a BP neural network using a genetic algorithm include population initialization, responsiveness function, selection operation, crossover operation, and mutation operation.

[0040] The hybrid dataset in step S5 is a combination of simulation data, self-developed experimental data, and literature data.

[0041] The beneficial effects of the CVD / CVI digital modeling and prediction method of this invention are as follows:

[0042] 1. This invention integrates CVD / CVI numerical simulation, database development, and machine learning-assisted prediction modeling. It combines the collected experimental data with data derived from numerical simulation to form a database, enabling the ML model to learn the "knowledge" of sparse regions in the literature experimental data and achieve higher-precision prediction in sparse regions.

[0043] 2. Due to the larger volume of simulated data compared to self-developed experimental data, the ML model can fully learn the reaction mechanism of CVD / CVI within this process parameter range during training, resulting in a significant improvement in prediction performance. The proposed model lays the theoretical and methodological foundation for future CVD / CVI digital twins and intelligent manufacturing. Attached Figure Description

[0044] Figure 1 This is a diagram of the SiC-CVD / CVI system of the present invention.

[0045] Figure 2 This is a geometric model diagram of the reactor of the present invention.

[0046] Figure 3 This is a grid model diagram of the reactor of the present invention.

[0047] Figure 4 This is a comparison chart of the simulated values ​​and the actual values ​​of the present invention.

[0048] Figure 5 This is a graph showing the prediction performance of the hybrid dataset ML, which is based on the fusion of simulated data, self-developed experimental data, and literature data, in the sparse region of the data.

[0049] Figure 6 This is a graph showing the prediction performance of the dataset ML, which is a fusion of self-developed experimental data and literature data from this invention, in the sparse region of the data. Detailed Implementation

[0050] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0051] Example 1

[0052] This invention provides a CVDCVI digital modeling and prediction method, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6As shown:

[0053] Step S1: A SiC coating or SiC substrate is prepared on a graphite substrate using the MTS(CH3SiCl3)-Ar-H2 system via chemical vapor deposition / chemical vapor infiltration (CVD / CVI). The CVD / CVI system used is as follows: Figure 1 As shown, the flow rates of H2 and Ar are controlled by the Mass Flow Controller, MFC.

[0054] The MTS solution is heated, and H2 bubbles in the MTS solution, carrying away the MTS. After all gases are thoroughly mixed in a mixing tank, they are piped into the chemical vapor deposition reactor. The reaction gases are gradually heated through a porous disc gas distributor, undergoing a gas-phase reaction during transport and adsorbing onto the graphite substrate to form SiC. The tail gas is discharged from the outlet.

[0055] Step S2: Due to the symmetry of the chemical vapor deposition reactor, a two-dimensional axisymmetric geometric model of the CVD / CVI reactor is established to reduce computational load. Figure 2 The geometric model is for a reactor with a reaction zone diameter of 100 mm. The reactor inlet diameter is 16 mm, and the length of the reactor reaction zone is either 200 mm or 400 mm. The front gas distribution plate is simplified as eight evenly arranged 3×3 mm rings.

[0056] The cylindrical structure inside the reactor is defined as a graphite substrate, serving as the surface for the chemical vapor deposition reaction. It has a diameter of 25 mm and a thickness of 6.5 mm. The MTS-H2-Ar mixed gas is introduced through the inlet on the left, evenly distributed through the gas distribution plate, and reacts inside the reactor to ultimately deposit SiC on the substrate surface. The exhaust gas is discharged through the exhaust port on the right side of the reactor.

[0057] Step S3: The two-dimensional axisymmetric geometric model of the chemical vapor deposition reactor was meshed using finite element software, and a mesh model of the reactor with a reaction zone diameter of 100 mm was constructed.

[0058] Step S3: In order to construct a simulation model of the SiC growth process, the following assumptions are made:

[0059] (1) Gas is a continuous substance;

[0060] (2) Gases obey the ideal gas law:

[0061] pV = nRT (1)

[0062] p is the pressure of an ideal gas, with units of Pa; V is the volume of an ideal gas, with units of m³. 3T is the thermodynamic temperature, in K; n is the amount of substance of the ideal gas, in mol; R is the ideal gas constant, 8.314 J / mol·K.

[0063] (3) The gas flows in a laminar flow within the reactor;

[0064] (4) The Fick diffusion model and additional convection transport mechanism are adopted;

[0065] (5) Ignore the effects of gravity;

[0066] (6) The dynamic viscosity and thermal conductivity of the mixture are neglected due to the influence of MTS, and only H2 and Ar are considered;

[0067] (7) The diffusion coefficient of each component is its diffusion coefficient in Ar;

[0068] (8) Neglecting radiant heat and reaction heat;

[0069] (9) The deposition reaction occurs only on the substrate surface.

[0070] Based on this, the conservation equations for mass, momentum, species, and energy of gas-phase heat and mass transport are given. Law of Conservation of Mass:

[0071]

[0072] Here, ρ is the mass density of the gas mixture in the ideal gas law, and u is the vector of average mass velocity.

[0073] Conservation of momentum:

[0074]

[0075] Here, ρ is the pressure on the fluid particle, μ is the dynamic viscosity, and I is the unit tensor.

[0076] Law of Conservation of Energy:

[0077]

[0078] Here, T is temperature, Cp is heat capacity, and k is thermal conductivity.

[0079] Conservation of matter:

[0080]

[0081] Here, ci is the molar concentration of the gas component, j is the diffusion flux, Ri is the reaction rate, and Di is the diffusion coefficient of the gas component.

[0082] The main gas-phase reaction of MTS(CH3SiCl3)-H2 is:

[0083] 1) The breaking of the Si-C bond in CH3SiCl3 initiates a decomposition reaction, generating CH3 and SiCl3 free radicals;

[0084] 2) These free radicals and their reactions with the original reacting gases (H2, CH3SiCl3) produce intermediate substances (such as alkanes like CH4, C2H2, C2H6 and silicon-based compounds like SiCl4, SiCl2, Si2Cl6) and byproducts (HCl), etc.

[0085] 3) Among them, intermediate substances with higher surface activity are more likely to be adsorbed onto the substrate surface, and then the adsorbed substances undergo surface reactions to generate SiC.

[0086] Therefore, the chemical vapor deposition simulation model constructed in this patent is mainly used for data derivation. Thus, using C2H4 and SiCl2 as the main active components, a three-step reaction is proposed to concisely describe the gas-phase-surface reaction of CVD / CVI-SiC. The kinetic constants of the reaction can be obtained through theoretical calculation. The chemical reactions and their kinetic constants used in this patent are shown in Table 1.

[0087] Table 1 Chemical reactions and their kinetic parameters

[0088] No Reaction <![CDATA[A(s -1 or m 3 mol -1 s -1 )]]> n <![CDATA[E(J / mol -1 )]]> G1 <![CDATA[CH3SiCl3→CH3+SiCl3]]> 7.63E14 0 2.9E5 G2 <![CDATA[2CH3→C2H4+H2]]> 8E11 0 8.5E4 G3 <![CDATA[SiCl3→SiCl2+H2]]> 3E9 0 2E5 G4 <![CDATA[SiCl2(s)+CH2(s)→SiC(s)+2HCl]]> 1E13 0 0 No Reaction τ <![CDATA[v a (1 / s)]]> Ea(eV) G5 <![CDATA[SiCl2→SiCl2(s)]]> 0.01 10^13 4 G6 <![CDATA[C2H4→2CH2(s)]]> Experimental Fit 10^13 3.5

[0089] Step S4: In Table 1, the A parameters of equation G6 need to be fitted based on existing experimental data. Combining this with the model, the "finding the optimal model parameter values" is transformed into a "function extremum optimization problem." Then, the data is input into a genetic algorithm. The elements of the genetic algorithm optimizing the BP neural network include population initialization, stress function, selection operation, crossover operation, and mutation operation, ultimately obtaining the A parameter values. The resulting model is then compared with the simulated results; the average simulation error is 12.8%.

[0090] Step S5: The performance of a machine learning model is affected not only by the amount of data in the training dataset, but also by the distribution of that data, which can significantly impact its overall prediction level.

[0091] Uneven data distribution, such as excessive concentration in a certain region, can cause the model to be biased towards a particular category, resulting in good predictive accuracy only within a certain range and poor overall predictive performance. To ensure the accuracy of subsequent digitization models and the uniformity of database distribution, the aforementioned model is used for simulation in sparse data areas to achieve data derivation and expand the database.

[0092] Table 2 shows the process parameters that need to be simulated for data derivation. The reactor with d=100mm underwent 45 simulations, generating a total of 45 sets of simulation data.

[0093] Process parameters derived from the data in Table 2

[0094]

[0095] Step S6: Establish a hybrid dataset, which is a combination of simulated data, self-developed experimental data, and literature data, used to train the ML model.

[0096] The self-developed experimental data can be newly measured offline data or newly measured experimental data through online sensors.

[0097] Online sensing can be achieved by connecting to a microbalance: the material is hung on one arm of the balance, and the weight change of the matrix during the deposition process is continuously recorded.

[0098] Step S7: For the mixed dataset, the XG Boost algorithm and the Random Forest algorithm can be used for training, and the results are as follows. Figure 5 Compared to datasets that combine purely self-developed experimental data and literature data (e.g.) Figure 6 As shown in the figure, the accuracy is higher. The results show that by establishing a numerical model of CVD deposition, expanding the data, and combining it with machine learning algorithms, the CVD deposition rate can be accurately predicted.

[0099] This invention is also an effective method for CVD processes in other multi-component systems.

[0100] Example 2

[0101] This invention provides a CVD / CVI digital modeling and prediction method, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown:

[0102] This invention is directed at CVD / CVI-SiC, but the proposed method can also be used for CVD / CVI material systems.

[0103] Example 3

[0104] This invention provides a CVD / CVI digital modeling and prediction method, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown:

[0105] Because the self-developed experimental data for CVD / CVI has the disadvantages of a wide range of process parameters and a small amount of data, the ML model cannot fully learn the reaction mechanism of CVD in this process parameter range, resulting in poor improvement in prediction performance.

[0106] This patent integrates CVD / CVI numerical simulation, database development, and machine learning-assisted predictive modeling. It combines the collected experimental data with data derived from numerical simulation to form a database, enabling the ML model to learn the "knowledge" of sparse regions in the literature experimental data and achieve higher-precision prediction in sparse regions.

[0107] Meanwhile, due to the larger volume of simulated data compared to self-developed experimental data, the ML model can fully learn the reaction mechanism of CVD / CVI within this process parameter range during training, resulting in a significant improvement in predictive performance. The proposed model lays the theoretical and methodological foundation for future CVD / CVI digital twins and intelligent manufacturing.

[0108] The protected points of this invention are:

[0109] 1. Establish a CVD / CVI simulation model, calibrate it with a small amount of experimental data, obtain the most reasonable model parameters using a genetic algorithm, and then use the CVD / CVI simulation model to generate a sufficient process simulation database, which greatly expands the data distribution and data volume in the database.

[0110] 2. Integrating CVD / CVI numerical simulation, database development, and machine learning-assisted predictive modeling, the collected experimental data is processed and merged with the data derived from numerical simulation to form a database, enabling the ML model to have higher accuracy in predicting sparse areas of experimental data.

Claims

1. A prediction method for a CVD / CVI digitization model, characterized in that, The CVD / CVI digitization model includes a first H2 bucket, an Ar bucket, and a second H2 bucket. 、 The system includes a chemical vapor deposition reactor, a gas mixer, and an electronic balance with heating function. The output ports of the first H2 tank, the Ar tank, and the electronic balance with heating function are all connected to the gas mixer. The output port of the second H2 tank is connected to the gas mixer, and the gas mixer is connected to the inlet of the chemical vapor deposition reactor. The chemical vapor deposition reactor has a graphite substrate on one side inside the reactor and a graphite heater on the other side inside the reactor. A porous disk is provided at the cross-section inside the chemical vapor deposition reactor and behind the inlet of the chemical vapor deposition reactor. An exhaust port is provided at the bottom cross-section of the chemical vapor deposition reactor. The inlet diameter of the chemical vapor deposition reactor is 16 mm, the length of the reaction zone of the chemical vapor deposition reactor is 200 mm or 400 mm, the diameter of the chemical vapor deposition reactor is 25 mm, and the thickness is 6.5 mm. The inlet diameter of the chemical vapor deposition reactor is 16 mm, and the length of the reaction zone of the chemical vapor deposition reactor includes 200 mm or 400 mm; the surface diameter of the chemical vapor deposition reactor is 25 mm, and the thickness is 6.5 mm. Includes the following steps: Step S1: Using the MTS(CH3SiCl3)-Ar-H2 system, SiC coating or SiC substrate is prepared on a graphite substrate by chemical vapor infiltration to generate SiC; Step S2: Establish a two-dimensional axisymmetric geometric model for the CVD / CVI chemical vapor deposition reactor; Step S3: Use finite element software to mesh the two-dimensional axisymmetric geometric model of the CVD / CVI chemical vapor deposition reactor, and construct a mesh model of the reactor with a reaction zone diameter of 100 mm, as well as the chemical reaction and its kinetic parameters. Step S4: Process the existing experimental data for parameter A in the chemical reaction and its kinetic parameters using an algorithm to finally obtain the value of parameter A; Step S5: Assess the A parameter values ​​and data distribution within the training dataset to form a mixed dataset; Step S6: Train the ML model based on the mixed dataset; Step S7: For the trained hybrid dataset, the XG Boost algorithm and the Random forest algorithm are used for training, and then combined with machine learning algorithms to predict the CVD / CVI deposition rate.

2. The prediction method for the CVD / CVI digital model according to claim 1, characterized in that, Step S1 includes the following sub-steps: Sub-step S11: The flow rates of H2 and Ar are controlled by a mass flow controller; Sub-step S12: By heating the MTS solution, H2 bubbles in the MTS solution and carries out the MTS. Sub-step S13: After thorough mixing in all gas mixers, the mixture is then piped into the chemical vapor deposition reactor; Sub-step S14: The reaction gas passes through a porous gas distribution plate and is gradually heated. During transport, a gas phase reaction occurs and SiC is adsorbed on the graphite substrate. The tail gas is discharged from the outlet.

3. The prediction method for the CVD / CVI digital model according to claim 1, characterized in that, The algorithm processing in step S4 includes the following sub-steps: Sub-step S41: The A parameter needs to be fitted based on the existing experimental data, and combined with the model, the "finding the optimal model parameter value" is transformed into a "function extremum optimization problem"; Sub-step S42: Input the genetic algorithm, which optimizes the elements of the BP neural network and finally obtains the A parameter value.

4. The prediction method for the CVD / CVI digital model according to claim 3, characterized in that, The elements of the genetic algorithm for optimizing the BP neural network include population initialization, responsiveness function, selection operation, crossover operation, and mutation operation.

5. The prediction method for the CVD / CVI digital model according to claim 1, characterized in that, The hybrid dataset in step S5 is a combination of simulation data, self-developed experimental data, and literature data.

Images